unique_toolkit 2026.22.0__tar.gz → 2026.22.0.dev0__tar.gz

{unique_toolkit-2026.22.0 → unique_toolkit-2026.22.0.dev0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: unique-toolkit
-Version: 2026.22.0
+Version: 2026.22.0.dev0
 Summary: 
 Author: Cedric Klinkert, Martin Fadler, Sadique Sheik, Fabian Schläpfer, Pascal Hauri, Ejiroghene Oki-Peter, Rami Azouz, Ricardo Visini, Sanzio Monti
 Author-email: Cedric Klinkert <cedric.klinkert@unique.ch>, Martin Fadler <martin.fadler@unique.ch>, Sadique Sheik <sadique@unique.ch>, Fabian Schläpfer <fabian@unique.ch>, Pascal Hauri <pascal@unique.ch>, Ejiroghene Oki-Peter <oki-peter.ejiroghene@unique.ai>, Rami Azouz <rami.ext@unique.ch>, Ricardo Visini <ricardo.visini@unique.ch>, Sanzio Monti <sanzio.monti@unique.ai>
@@ -16,7 +16,7 @@ Requires-Dist: sseclient>=0.0.27,<0.0.28
 Requires-Dist: openai>=1.105.0,<3
 Requires-Dist: platformdirs>=4.0.0,<5
 Requires-Dist: pillow>=12.1.1
-Requires-Dist: unique-sdk>=2026.22.0
+Requires-Dist: unique-sdk>=2026.22.0.dev0,<2026.22.0rc0
 Requires-Dist: jambo>=0.1.2,<0.2
 Requires-Dist: jinja2>=3.1.6,<4
 Requires-Dist: docxtpl>=0.20.1,<0.21

{unique_toolkit-2026.22.0 → unique_toolkit-2026.22.0.dev0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "unique_toolkit"
-version = "2026.22.0"
+version = "2026.22.0.dev0"
 description = ""
 readme = "README.md"
 requires-python = ">=3.12"
@@ -28,7 +28,7 @@ dependencies = [
     "openai>=1.105.0,<3",
     "platformdirs>=4.0.0,<5",
     "pillow>=12.1.1",
-    "unique-sdk>=2026.22.0",
+    "unique-sdk>=2026.22.0.dev0,<2026.22.0rc0",
     "jambo>=0.1.2,<0.2",
     "jinja2>=3.1.6,<4",
     "docxtpl>=0.20.1,<0.21",

{unique_toolkit-2026.22.0 → unique_toolkit-2026.22.0.dev0}/unique_toolkit/_common/utils/files.py

@@ -8,7 +8,6 @@ from pathlib import Path
 class FileMimeType(StrEnum):
     PDF = "application/pdf"
     DOCX = "application/vnd.openxmlformats-officedocument.wordprocessingml.document"
-    DOTX = "application/vnd.openxmlformats-officedocument.wordprocessingml.template"
     DOC = "application/msword"
     # TODO: clean up duplicates and make the monolith compatible with this.
     XLSX = "application/vnd.openxmlformats-officedocument.spreadsheetml.sheet"
@@ -42,7 +41,7 @@ class FileMimeType(StrEnum):
     @property
     def is_docx(self) -> bool:
         cls = self.__class__
-        return self in {cls.DOCX, cls.DOTX, cls.DOC}
+        return self in {cls.DOCX, cls.DOC}
 
     @property
     def is_pdf(self) -> bool:
@@ -99,7 +98,7 @@ class FileMimeType(StrEnum):
 
 def get_common_name(extension: FileMimeType) -> str:
     match extension:
-        case FileMimeType.DOCX | FileMimeType.DOTX | FileMimeType.DOC:
+        case FileMimeType.DOCX | FileMimeType.DOC:
             return "docx"
         case (
             FileMimeType.XLSX

{unique_toolkit-2026.22.0 → unique_toolkit-2026.22.0.dev0}/unique_toolkit/agentic/feature_flags/feature_flags.py

@@ -66,6 +66,11 @@ class FeatureFlags(BaseSettings):
         description="Enable new answers UI (UN-14411)",
     )
 
+    enable_elicitation_un_15809: FeatureFlag = Field(
+        default=FeatureFlag(False),
+        description="Enable elicitation (UN-15809)",
+    )
+
     enable_html_rendering_un_15131: FeatureFlag = Field(
         default=FeatureFlag(False),
         description="Enable HTML rendering for code interpreter files (UN-15131). Can be 'true' or comma-separated company IDs.",
@@ -76,11 +81,6 @@ class FeatureFlags(BaseSettings):
         description="Emit codeExecution fences in message.text for code interpreter outputs (UN-17972). When disabled, inline file refs are kept as-is. Can be 'true' or comma-separated company IDs.",
     )
 
-    enable_html_with_fence_un_17927: FeatureFlag = Field(
-        default=FeatureFlag(False),
-        description="Render HTML code interpreter files as htmlWithSource fences instead of HtmlRendering blocks (UN-17927). Requires enable_code_execution_fence_un_17972 to also be enabled. Can be 'true' or comma-separated company IDs.",
-    )
-
     enable_web_search_argument_screening_un_18741: FeatureFlag = Field(
         default=FeatureFlag(False),
         description="Enable argument screening agent for web search tool calls (UN-18741). Can be 'true' or comma-separated company IDs.",

{unique_toolkit-2026.22.0 → unique_toolkit-2026.22.0.dev0}/unique_toolkit/agentic/history_manager/loop_token_reducer.py

@@ -13,7 +13,6 @@ from unique_toolkit.agentic.history_manager.history_construction_with_contents i
     get_full_history_with_contents_and_tool_calls_async,
     get_full_history_with_contents_async,
 )
-from unique_toolkit.agentic.history_manager.settings import env_settings
 from unique_toolkit.agentic.history_manager.utils import (
     get_selected_uploaded_content_ids,
     serialize_tool_content_json,
@@ -34,8 +33,10 @@ from unique_toolkit.language_model.schemas import (
 )
 from unique_toolkit.services.factory import UniqueServiceFactory
 
-MAX_INPUT_TOKENS_SAFETY_PERCENTAGE = env_settings.input_correction_reduction_factor
-SAFETY_MARGIN_FOR_AGGRESSIVE_REDUCTION = env_settings.input_aggressive_reduction_factor
+MAX_INPUT_TOKENS_SAFETY_PERCENTAGE = (
+    0.1  # 10% safety margin for input tokens we need 10% less does not work.
+)
+SAFETY_MARGIN_FOR_AGGRESSIVE_REDUCTION = 0.9
 
 
 class SourceReductionResult(BaseModel):

{unique_toolkit-2026.22.0 → unique_toolkit-2026.22.0.dev0}/unique_toolkit/agentic/tools/config.py

@@ -5,8 +5,8 @@ from pydantic import (
     BaseModel,
     BeforeValidator,
     Field,
-    SerializeAsAny,
     ValidationInfo,
+    field_serializer,
     model_validator,
 )
 from typing_extensions import TypeVar
@@ -53,7 +53,7 @@ class ToolBuildConfig(BaseModel, Generic[T]):
     """Main tool configuration"""
 
     name: str
-    configuration: SerializeAsAny[T]
+    configuration: T
     display_name: str = ""
     icon: Annotated[ToolIcon, BeforeValidator(handle_undefined_icon)] = Field(
         default=ToolIcon.BOOK,
@@ -125,3 +125,7 @@ class ToolBuildConfig(BaseModel, Generic[T]):
             config = configuration
         value["configuration"] = config
         return value
+
+    @field_serializer("configuration")
+    def serialize_config(self, value: T) -> dict[str, Any]:
+        return value.__class__.model_dump(value)

{unique_toolkit-2026.22.0 → unique_toolkit-2026.22.0.dev0}/unique_toolkit/agentic/tools/names.py

@@ -8,4 +8,3 @@ TODO(UN-20578): Add canonical name constants for the remaining toolkit tools.
 
 INTERNAL_SEARCH_TOOL_NAME = "InternalSearch"
 UPLOADED_SEARCH_TOOL_NAME = "UploadedSearch"
-SKILL_TOOL_NAME = "Skill"

{unique_toolkit-2026.22.0 → unique_toolkit-2026.22.0.dev0}/unique_toolkit/agentic/tools/openai_builtin/base.py

@@ -46,13 +46,6 @@ class OpenAIBuiltInTool(ABC, Generic[ToolType]):
     def display_name(self) -> str:
         raise NotImplementedError()
 
-    def is_capability(self) -> bool:
-        """Return True if this tool is a capability that must remain in the
-        active tool set regardless of ``tool_choices`` filtering.
-        Default: False.
-        """
-        return False
-
     def get_required_include_params(self) -> list[ResponseIncludable]:
         """Return Responses API `include` values required by this tool.
 

{unique_toolkit-2026.22.0 → unique_toolkit-2026.22.0.dev0}/unique_toolkit/agentic/tools/openai_builtin/code_interpreter/config.py

@@ -17,36 +17,7 @@ from unique_toolkit.agentic.tools.openai_builtin.code_interpreter.postprocessors
 )
 from unique_toolkit.agentic.tools.schemas import BaseToolConfig
 
-_TOOL_DESCRIPTION_FILES_SUFFIX = """
-{%- if user_uploaded_files and user_uploaded_files | length > 0 %}
-
-## User Uploaded Files
-The following files have been uploaded by the user and are available in the environment:
-{% for file in user_uploaded_files %}
-- {{ file }}
-{% endfor %}
-{% endif %}
-
-{%- if kb_uploaded_files and kb_uploaded_files | length > 0 %}
-
-## Knowledge Base Files
-These files are made available by the system but were NOT uploaded by the user:
-{% for file in kb_uploaded_files %}
-- {{ file }}
-{% endfor %}
-{% endif %}
-
-{%- if code_interpreter_artifacts and code_interpreter_artifacts | length > 0 %}
-
-## Code Interpreter Artifacts
-The following files have been generated by you using the python tool in previous turns:
-{% for file in code_interpreter_artifacts %}
-- {{ file }}
-{% endfor %}
-{% endif %}
-""".strip()
-
-_TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT_BASE = """
+DEFAULT_TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT = """
 Use this tool to run python code, e.g to generate plots, process excel files, perform calculations, etc.
 
 Instructions:
@@ -66,7 +37,7 @@ WHENEVER you reference a generated file, you MUST use the following format:
 IMPORTANT: Do NOT append a leading `!` even when displaying an image.
 Always use a line break between the title and the markdown!
 - Files with image file extensions are displayed directly in the chat, while other file extensions are shown as download links.
-- Not using syntax above will FAIL to show images to the user.
+- Not using syntax above will FAIL to show images to the user. 
 - YOU MUST use the syntax above to display images, otherwise the image will not be displayed in the chat.
 - Only the following file types are allowed to be uploaded to the platform, anything else will FAIL: PDF, DOCX, XLSX, PPTX, CSV, HTML, MD, TXT, PNG, JPG, JPEG.
 - You MUST always use this syntax, otherwise the files will not be displayed in the chat.
@@ -87,11 +58,7 @@ Handling User Queries:
         - REMEMBER that you can always read the content of text, csv files if needed. In this case, you MUST always limit the amount of data displayed.
 - If a tool step fails, you must recover as much as possible.
 - After exhausting all possible solutions without success, inform the user of what was tried and request clarification/help.
-"""
-
-DEFAULT_TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT = (
-    _TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT_BASE + _TOOL_DESCRIPTION_FILES_SUFFIX
-).strip()
+""".strip()
 
 # Used when the code-execution fence feature flag (UN-17972) is enabled.
 # The frontend derives the artifact title from the filename itself, so the
@@ -104,7 +71,7 @@ DEFAULT_TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT = (
 # UN-19880: kaleido is not installed in the Azure OpenAI code interpreter environment,
 # so plotly's fig.write_image() always fails. The "Visualization library choice" section
 # steers the model to use matplotlib for PNG outputs and plotly only for HTML artifacts.
-_TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT_FENCE_BASE = """
+DEFAULT_TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT_FENCE = """
 Use this tool to run python code, e.g to generate plots, process excel files, perform calculations, etc.
 
 Instructions:
@@ -170,11 +137,7 @@ Handling User Queries:
         - REMEMBER that you can always read the content of text, csv files if needed. In this case, you MUST always limit the amount of data displayed.
 - If a tool step fails, you must recover as much as possible.
 - After exhausting all possible solutions without success, inform the user of what was tried and request clarification/help.
-"""
-
-DEFAULT_TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT_FENCE = (
-    _TOOL_DESCRIPTION_FOR_SYSTEM_PROMPT_FENCE_BASE + _TOOL_DESCRIPTION_FILES_SUFFIX
-).strip()
+""".strip()
 
 
 DEFAULT_TOOL_DESCRIPTION_FOR_USER_PROMPT = ""

{unique_toolkit-2026.22.0 → unique_toolkit-2026.22.0.dev0}/unique_toolkit/agentic/tools/openai_builtin/code_interpreter/postprocessors/generated_files.py

@@ -21,6 +21,7 @@ from tenacity import (
 
 from unique_toolkit import ChatService
 from unique_toolkit._common.execution import failsafe_async
+from unique_toolkit._common.utils.files import FileMimeType, ImageMimeType
 from unique_toolkit.agentic.feature_flags.feature_flags import feature_flags
 from unique_toolkit.agentic.postprocessor.postprocessor_manager import (
     ResponsesApiPostprocessor,
@@ -29,9 +30,6 @@ from unique_toolkit.agentic.short_term_memory_manager.persistent_short_term_memo
     PersistentShortMemoryManager,
 )
 from unique_toolkit.agentic.tools.config import get_configuration_dict
-from unique_toolkit.agentic.tools.openai_builtin.code_interpreter.postprocessors.artifacts import (
-    save_code_execution_artifact,
-)
 from unique_toolkit.agentic.tools.openai_builtin.code_interpreter.schemas import (
     CodeInterpreterBlock,
     CodeInterpreterFile,
@@ -503,15 +501,6 @@ class DisplayCodeInterpreterFilesPostProcessor(
         fence_ff_on = feature_flags.enable_code_execution_fence_un_17972.is_enabled(
             self._company_id
         )
-        # HTML uses htmlWithSource only when BOTH the code-execution fence FF AND the
-        # dedicated HTML-fence FF are on. Default (FF off) keeps HtmlRendering so
-        # existing deployments are unaffected.
-        html_fence_ff_on = (
-            fence_ff_on
-            and feature_flags.enable_html_with_fence_un_17927.is_enabled(
-                self._company_id
-            )
-        )
 
         replaced_files: list[str] = []
         missed_files: list[str] = []
@@ -543,9 +532,7 @@ class DisplayCodeInterpreterFilesPostProcessor(
                 )
                 changed |= replaced
 
-            # HTML rendered as HtmlRendering block unless the dedicated html-fence FF
-            # is on (enable_html_with_fence_un_17927). Default keeps HtmlRendering.
-            elif is_html and not html_fence_ff_on:
+            elif is_html:
                 loop_response.message.text, replaced = _replace_container_html_citation(
                     text=loop_response.message.text or "",
                     filename=filename,
@@ -553,8 +540,6 @@ class DisplayCodeInterpreterFilesPostProcessor(
                 )
                 changed |= replaced
 
-            # Non-HTML files, or HTML when html_fence_ff_on → inline content link for
-            # subsequent fence injection (htmlWithSource / imgWithSource / fileWithSource)
             else:
                 loop_response.message.text, replaced = _replace_container_file_citation(
                     text=loop_response.message.text or "",
@@ -576,7 +561,6 @@ class DisplayCodeInterpreterFilesPostProcessor(
             else:
                 missed_files.append(filename)
 
-            # HtmlRendering and htmlWithSource both embed contentId directly — no ContentReference needed
             is_html_rendered = is_html
             if replaced and not (is_image or is_html_rendered or fence_ff_on):
                 loop_response.message.references.append(
@@ -598,17 +582,13 @@ class DisplayCodeInterpreterFilesPostProcessor(
         )
 
         if fence_ff_on:
-            code_blocks = _build_code_blocks(
-                loop_response, self._content_map, include_html=html_fence_ff_on
-            )
+            code_blocks = _build_code_blocks(loop_response, self._content_map)
             self._log.info(
                 "Fence injection — %d code block(s), files: %s",
                 len(code_blocks),
                 [f.filename for b in code_blocks for f in b.files],
             )
-            _warn_unmatched_code_blocks(
-                self._content_map, code_blocks, include_html=html_fence_ff_on
-            )
+            _warn_unmatched_code_blocks(self._content_map, code_blocks)
             text_before = loop_response.message.text
             loop_response.message.text = _inject_code_execution_fences(
                 loop_response.message.text or "",
@@ -681,13 +661,32 @@ class DisplayCodeInterpreterFilesPostProcessor(
                     container_file.filename, "uploading", force_publish=True
                 )
 
+            raw_mime = guess_type(container_file.filename)[0] or "text/plain"
+            mime = _kb_safe_mime(raw_mime)
+            if mime != raw_mime:
+                self._log.info(
+                    "MIME type '%s' is not supported by the Unique KB; "
+                    "uploading '%s' as '%s' so the file can be stored and downloaded.",
+                    raw_mime,
+                    container_file.filename,
+                    mime,
+                )
+            self._log.info(
+                "Uploading '%s' to knowledge base (mime=%s, %d bytes)",
+                container_file.filename,
+                mime,
+                len(file_bytes),
+            )
+
             assert self._chat_service is not None  # Checked in __init__
             upload_t0 = time.monotonic()
             content = await self._build_retry()(
-                save_code_execution_artifact,
-                chat_service=self._chat_service,
-                file=container_file,
-                file_bytes=file_bytes,
+                self._chat_service.upload_to_chat_from_bytes_async,
+                content=file_bytes,
+                content_name=container_file.filename,
+                mime_type=mime,
+                skip_ingestion=True,
+                hide_in_chat=True,
             )
             upload_ms = (time.monotonic() - upload_t0) * 1000
             if upload_ms > 2000:
@@ -930,6 +929,37 @@ def _get_file_type(filename: str) -> CodeInterpreterFileType:
     return "document"
 
 
+def _kb_safe_mime(mime: str) -> str:
+    """Return a MIME type the Unique KB will accept.
+
+    Membership is defined by ``FileMimeType`` and ``ImageMimeType`` in
+    ``unique_toolkit._common.utils.files`` (same catalog as path-based helpers
+    like ``FileMimeType.is_valid_mime``, but here we already have a resolved
+    MIME string from ``mimetypes.guess_type``, so we use StrEnum value lookup
+    instead of re-parsing a path).
+
+    The KB GraphQL API rejects many code-file MIME types (e.g. ``text/x-python``
+    for ``.py`` files).  Anything not in those enums is coerced to
+    ``text/plain`` so the file can be stored and downloaded without changing
+    its bytes.
+
+    Other ``image/*`` subtypes (e.g. ``image/jpg``) still pass through unchanged.
+    """
+    try:
+        FileMimeType(mime)
+        return mime
+    except ValueError:
+        pass
+    try:
+        ImageMimeType(mime)
+        return mime
+    except ValueError:
+        pass
+    if mime.startswith("image/"):
+        return mime
+    return "text/plain"
+
+
 def _file_title(filename: str) -> str:
     """Derive a human-readable title from a filename.
 
@@ -989,9 +1019,12 @@ def _build_file_fence(file: CodeInterpreterFile, code: str, fence_id: int) -> st
     content_id, a title derived from the filename, the type, and the
     escaped source code.
 
+    HTML is not emitted here in normal flow: it is rendered via ``HtmlRendering`` blocks
+    in message text and excluded from fence injection. If ``type="html"`` is passed (e.g.
+    orphan path), fall through to ``fileWithSource`` like other non-image artifacts.
+
     Format (4 backticks so inner code backticks never close the fence):
       ````imgWithSource(id='{n}', contentId='{id}', title="{title}", code="{escaped_code}")````
-      ````htmlWithSource(id='{n}', contentId='{id}', title="{title}", code="{escaped_code}")````
       ````fileWithSource(id='{n}', contentId='{id}', title="{title}", type="{type}", code="{escaped_code}")````
     """
     title = _file_title(file.filename)
@@ -999,8 +1032,6 @@ def _build_file_fence(file: CodeInterpreterFile, code: str, fence_id: int) -> st
     outer = "````"
     if file.type == "image":
         tag = f"imgWithSource(id='{fence_id}', contentId='{file.content_id}', title=\"{title}\", code=\"{escaped}\")"
-    elif file.type == "html":
-        tag = f"htmlWithSource(id='{fence_id}', contentId='{file.content_id}', title=\"{title}\", code=\"{escaped}\")"
     else:
         ftype = _file_frontend_type(file.filename)
         tag = f'fileWithSource(id=\'{fence_id}\', contentId=\'{file.content_id}\', title="{title}", type="{ftype}", code="{escaped}")'
@@ -1024,7 +1055,7 @@ def _inline_ref_pattern(file: CodeInterpreterFile) -> re.Pattern[str]:
 
 
 _FENCE_BLOCK_START = re.compile(
-    r"^[^\n`]+?(````(?:imgWithSource|fileWithSource|htmlWithSource)\()",
+    r"^[^\n`]+?(````(?:imgWithSource|fileWithSource)\()",
     re.MULTILINE,
 )
 
@@ -1034,9 +1065,9 @@ _FENCE_BLOCK_START = re.compile(
 #   - cross-line: 1 or 2 newlines optionally surrounded by horizontal whitespace
 #     (list-item linebreak, or blank-line paragraph gap)
 _CONSECUTIVE_FENCES_RE = re.compile(
-    r"(````(?:imgWithSource|fileWithSource|htmlWithSource)\([^\n]*\)````)"
+    r"(````(?:imgWithSource|fileWithSource)\([^\n]*\)````)"
     r"(?:[^\n`]*|[ \t]*\n{1,2}[ \t]*)"
-    r"(?=````(?:imgWithSource|fileWithSource|htmlWithSource)\()"
+    r"(?=````(?:imgWithSource|fileWithSource)\()"
 )
 
 
@@ -1100,7 +1131,6 @@ def _inject_code_execution_fences(
 def _build_code_blocks(
     loop_response: ResponsesLanguageModelStreamResponse,
     content_map: dict[str, str | None],
-    include_html: bool = False,
 ) -> list[CodeInterpreterBlock]:
     """Map each code interpreter call to the files it produced via /mnt/data/ path matching.
 
@@ -1198,9 +1228,11 @@ def _build_code_blocks(
     # overwritten across executions. Using a dict keyed by filename ensures each
     # file appears exactly once per block (last annotation wins, consistent with
     # the last-writer-wins rule applied in step 1).
+    # HTML files are excluded: they are always rendered via HtmlRendering blocks
+    # and never participate in fence injection.
     block_file_map: dict[int, dict[str, CodeInterpreterFile]] = {}
     for annotation in loop_response.container_files:
-        if not include_html and _get_file_type(annotation.filename) == "html":
+        if _get_file_type(annotation.filename) == "html":
             continue
         content_id = content_map.get(annotation.filename)
         if content_id is None:
@@ -1284,7 +1316,6 @@ def _replace_dangling_sandbox_links(text: str) -> tuple[str, bool]:
 def _warn_unmatched_code_blocks(
     content_map: dict[str, str | None],
     code_blocks: list[CodeInterpreterBlock],
-    include_html: bool = False,
 ) -> None:
     """Warn for files that were uploaded but could not be matched to any code block.
 
@@ -1298,7 +1329,7 @@ def _warn_unmatched_code_blocks(
     for filename, content_id in content_map.items():
         if content_id is None:
             continue
-        if not include_html and _get_file_type(filename) == "html":
+        if _get_file_type(filename) == "html":
             continue
         if filename not in fenced_filenames:
             logger.warning(